Objectives

Can we issue a buy/sell recommendation of cryptostock on a 7days holding period?

What is the relevant Data for our model ?

Price
Volume
Correlation
Social Media
Google Trends
Utility Indicator

Data Preparation

We used an external Python script to scrape the website www.coinmarketcap.com

Besides technical analysis on the stock, we also wanted to include a trend factor with our data so we looked at Google Trend.

Our hypothesis was that there is a strong correlation between google searches and stock prices.

Feature Overview

Volume
Momemtum
Volatility
Trend
Buy / Sell classifier

Downloading Data

After scrapping coinmarketcap website, we had the following data:

We also managed to download google data using a specific R library.

scrapGTrendsForKeywords(c("BTC","ETH","XRP","EOS","LTC"), "gtrends.csv")

Atht he moment we are unable to sort our currency by market cap.

Cleanup data

Remove unused IDs
Format Dates
Interpolate missing data
- Locate missing dates, insert row, and interpolate values

Visualize initial Data

plotGTrends(google.trends)

Feature Engineering

We build these categorical variables on different lag periods:

7 days
14 days
21 days

Create Overall Market statistics

The initial part of the analysis is to be able to create a new dataframe that will contain the total history of the market and useful indicators for technical analysis.

Total market Cap daily
Volatility 7d, 30d, 90d
returns + logreturns
Volume

##  [1]  1  2  3  4  5  6  7  8  9 10 11 12 13 14 15 16 17 18 19 20 21 22 23
## [24] 24 25 26 27


name	type	slug	mcap	beta
BTC	coin	bitcoin	110701963288	0.9773526
ETH	coin	ethereum	49547821373	0.9059106
XRP	coin	ripple	20953892318	0.9308553
BCH	coin	bitcoin-cash	14724786964	1.0251973
EOS	token	eos	9552774356	1.2269009
LTC	coin	litecoin	5498122696	0.9448140
XLM	coin	stellar	4303693082	0.9678134
ADA	coin	cardano	4225775445	1.4754460
MIOTA	coin	iota	3291464170	1.0917421
TRX	token	tron	2829634354	1.4993969
USDT	token	tether	2618194620	-0.0132857
NEO	coin	neo	2508226500	0.8396604
DASH	coin	dash	2094727525	0.6616042
XMR	coin	monero	1989915245	0.9092103
XEM	coin	nem	1750283999	1.0847094
BNB	token	binance-coin	1732275792	1.1453165
VEN	token	vechain	1571361432	1.1307183
ETC	coin	ethereum-classic	1431992689	0.9812628
QTUM	coin	qtum	948949715	0.8378323
OMG	token	omisego	932475042	0.9637187
ONT	token	ontology	907303711	0.6437450
ZEC	coin	zcash	798735715	0.8147235
BCN	coin	bytecoin-bcn	774143986	0.7442515
ICX	token	icon	763252332	1.0782522
LSK	coin	lisk	703390829	0.8420246
ZIL	token	zilliqa	661689897	1.2276519
DCR	coin	decred	650912867	0.6353398

“Scaled” Google Trends

Google only provide daily data up to 3months (automatically switch to weekly above 3months range)
Data provided is “scaled” relative to the period (meaning 1-100 on each period instead of absolute value)

plotGTrendsIssue(google.trends)

Volume

Represent the average volume for a period:

volume.7d
volume.14d
volume.21d

Volatility

Standard deviation of the returns for a period:

volatility.7d
volatility.14d
volatility.21d

Momemtum

Score representing the delta of Upwards VS Downwards returns.

momentum.7d
momentum.14d
momemtum.21d

Trends

Linear regression of the Trends for a period from which we extract the coeeficient.

trends.7d
trends.14d
trends.21d

Buy / Sell Classifier

Starting from a specific day, we look back at a period returns to determine if we should have issued a buy or sell. This will allow us to test our model.

buy.7d
buy.14d
buy.21d

Putting it together

We are able to define the state of the market by mixing different features together. For example, we can determine how bullish/bearish a market is by multiplying our trend coefficient with the momentum.

btcValues = coinDataEngineering("BTC")

## [BTC] Feature engineering on full dataset for
##  [BTC] SMA20 [Done]
##  [BTC] Volatility [Done]
##  [BTC] Volume [Done]
##  [BTC] Momemtum [Done]
##  [BTC] BuyResult [Done]
##  [BTC] GoogleTrends [Done]

ethValues = coinDataEngineering("ETH")

## [ETH] Feature engineering on full dataset for
##  [ETH] SMA20 [Done]
##  [ETH] Volatility [Done]
##  [ETH] Volume [Done]
##  [ETH] Momemtum [Done]
##  [ETH] BuyResult [Done]
##  [ETH] GoogleTrends [Done]

xrpValues = coinDataEngineering("XRP")

## [XRP] Feature engineering on full dataset for
##  [XRP] SMA20 [Done]
##  [XRP] Volatility [Done]
##  [XRP] Volume [Done]
##  [XRP] Momemtum [Done]
##  [XRP] BuyResult [Done]
##  [XRP] GoogleTrends [Done]

ltcValues = coinDataEngineering("LTC")

## [LTC] Feature engineering on full dataset for
##  [LTC] SMA20 [Done]
##  [LTC] Volatility [Done]
##  [LTC] Volume [Done]
##  [LTC] Momemtum [Done]
##  [LTC] BuyResult [Done]
##  [LTC] GoogleTrends [Done]

eosValues = coinDataEngineering("EOS")

## [EOS] Feature engineering on full dataset for
##  [EOS] SMA20 [Done]
##  [EOS] Volatility [Done]
##  [EOS] Volume [Done]
##  [EOS] Momemtum [Done]
##  [EOS] BuyResult [Done]
##  [EOS] GoogleTrends [Done]

Techincal Analysis

Logistic Regression

buy.7 ~ volume.7 + volume.14 + volume.21 + volatility.7 + volatility.14 + volatility.21 + momentum.7 + momentum.14 + momentum.21 + gtrend.7 + gtrend.14 + gtrend.21

model = buy.7 ~ volume.7 + volume.14 + volume.21 + volatility.7 + volatility.14 + volatility.21 + momentum.7 +  momentum.14 + momentum.21 + gtrend.7 + gtrend.14 + gtrend.21
  # Train our model first
  logistic_reg = glm(model, data=training, family="binomial"(link="logit"))
  
btcResults = doLogisticReg(btcValues)
ethResults = doLogisticReg(ethValues)
xrpResults = doLogisticReg(xrpValues)
ltcResults = doLogisticReg(ltcValues)
eosResults = doLogisticReg(eosValues)

Bitcoin

Ethereum

Ripple

EOS

Litecoin

Summary Results


Name	Accuracy	Sensitivities	Specificities	AUC
bitcoin	0.561538461538462	0.6	0.523076923076923	0.5760947
ethereum	0.449612403100775	0.5	0.39344262295082	0.4975892
ripple	0.705426356589147	0.734177215189873	0.66	0.7253165
litecoin	0.565891472868217	0.602564102564103	0.509803921568627	0.5892408
eos	0.75	0.6	0.9	0.7700000

Going Further

We have built a solid base to complete a better analysis in the future. Here is a list of topics we can investigate building on our current status:

Use probability to have categorized recommendation (Strong buy, Strong sell, neutral, …)
Portfolio Management / Optimization
- Instead of choosing top 5 by market cap, analysis can be updated daily/hourly on all top currencies
- Portfolio rebalancing
Identitfy arbitrage opportunities
Dynamic Horizons (already ready for 7,14,21 days)

Crypto Stock Analysis